Text analysis: title and abstract of male and female speakers

Data

Data description and summary statistics in script 0_data_summary.

load("data_clean/data_title.Rdata")
data <- data_t

Using data from titles and abstracts.

Formating titles

tit <- data %>% dplyr::select(id,gender,position_cat, audience_n,
                              title_english) 
text_tit <- tit %>% unnest_tokens(output=word,
                                   input=title_english)

Total of 320 titles.

Number of titles per group

table(tit$gender) %>% kable()

Var1	Freq
F	140
M	180

table(tit$position_cat,tit$gender) %>% kable()

	F	M
others	4	1
postdoc	25	32
professor	24	72
student	86	74

Formating abstracts

data_abs <- data  %>% filter(!is.na(abstract_english)) 
text_tok <- data %>% dplyr::select(id,gender,position_cat, audience_n,
                             abstract_english) %>%
  unnest_tokens(output=word,input=abstract_english)

Total of 234 abstracts.

Number of abstracts per group

table(data_abs$gender) %>% kable()

Var1	Freq
F	99
M	135

table(data_abs$position_cat,data_abs$gender)%>% kable()

	F	M
others	4	1
postdoc	21	21
professor	20	57
student	53	56

Tidytext

texts <- bind_rows(text_tit, text_tok)

stop_w <- tibble(word = stopwords("en"))

# remove stopwords
text <- texts %>% 
  anti_join(stop_w, by="word") %>% arrange(word) 

# remove other non-words (numbersm characters) and stopwords
text <- text %>% dplyr::slice(-c(1:281)) %>% # number and some symbols
          filter(nchar(word)!=1) %>% # letters alone
          filter(!word %in% c("mpas", "δ13c", "β", "can", "aff", "agb", "al",
                              "and","são"))

# solving some simple plurals
plural <- c("actions","advances", "adaptations", "amphibians", "animals", 
            "ants","anurans",
            "applications","approaches", "bees","builds", "birds",
            "cerrados","challenges", "biologists", "captures",
            "continents","crops", 
            "decisions","declines","determines","determinants", "defenses",
            "dynamics", "dunnocks",
            "economics", "ecosystems","environments", "experiences",
            "forests", "fruits", "fathers",
            "genetics","gifts","gradients","guides","impacts", "islands",
            "increases","interactions", "jobs",  "lives", "insects",
            "landscapes","males","mammals", "mangroves","models","movements",
            "mutualisms","networks","neotropics",
            "opilions","phenotypes","plants","projects","paths", "perspectives",
            "purposes", "populations","promotes","relationships", "relations",
            "resources","responses","roads","services","skulls","snakes","seeds",
            "soils", "spaces", "spiders","stages", "trees", "variations",
            "threats")

text$word[text$word %in% plural] <- 
  substr(text$word[text$word %in% plural],
       1,nchar(text$word[text$word %in% plural])-1)

Grouping similar words:

lemma <- rbind(c("adaptive", "adaptation"),
               c("advancement", "advance"),
               c("agricultural", "agriculture"),
               c("agro", "agriculture" ),
               c("amazonia","amazon" ),
               c("amazonian","amazon" ),
               c("andean","andes"),
               c("apply","application"),
               c("applying","application"),
               c("approaches", "approach"),
               c("apidae","apis"),
               c("arachnida","arachnid"),
               c("argue","argument"),
               c("basal", "basis"),
               c("behavioral","behavior"),
               c("behavioural","behavior"),
               c("bignonieae", "bignoniaceae"),
               c("biological", "biology"),
               c("brazilian","brazil"),
               c("building","build"),
               c("butterflies", "butterfly"),
               c("changing", "change"),
               c("cnidarian", "cnidaria"),
               c("coastal","coast"),
               c("colour", "color"),
               c("colors", "color"),
               c("communities","community" ),
               c("competitive", "competition"),
               c("complexity", "complex"),
               c("convergences", "convergence"),
               c("convergent", "convergence"),
               c("croplands","crop"),
               c( "cultural", "culture"),
               c("darwin's", "darwin"),
               c("darwinian", "darwin"),
               c("defensive", "defense"),
               c("dependent","dependence"),
               c("detecting","detection"),
               c("determine", "determinant"),
               c("developmental", "development"),
               c("dispersers","dispersal"),
               c("disturbed", "disturbance"),
               c("diversification", "diversity"),
               c("dragonflies", "dragonfly"),
               c("drier", "drought"),
               c("ecological", "ecology"),
               c("ecologists", "ecology"),
               c("endemic", "endemism"),
               c("effectiveness", "efficiency"),
               c("environmental", "environment"),
               c("evolutionary", "evolution"),
               c("expanding", "expansion"),
               c("extinct", "extinction"),
               c("facilitate", "facilitation"),
               c("fisheries", "fishery"),
               c("floral", "flora"),
               c("floristic", "flora"),
               c("forested", "forest"),
               c("functional", "function"),
               c("functionally", "function"),
               c("functioning", "function"),
               c("geographical", "geographic"),
               c("heterogeneties", "heterogeneity"),
               c("heterogeneous", "heterogeneity"),
               c("histories", "history"),
               c("integrated", "integration"),
               c("intregating", "integration"),
               c("integrative", "integration"),
               c("invasive", "invasion"),
               c("isotopic", "isotope"),
               c("linking", "link"),
               c("living", "live"),
               c("mammalia", "mammal"),
               c("managed", "manage"),
               c("managers", "manage"),
               c("mathematical", "mathematics"),
               c("mates", "mating"),
               c("mediated", "mediate"),
               c("mechanistic", "mechanism"),
               c("matrices", "matrix"),
               c("migratory", "migration"),
               c("mimicking", "mimicry"),
               c("modeling", "model"),
               c("mutualistic", "mutualism"),
               c("natural", "nature"),
               c("neotropical", "neotropic"),
               c("northeastern", "northeast"),
               c("occuring", "occur"),
               c("onça", "onca"),
               c("opiliones", "opilion"),
               c("parasite", "parasitism"),
               c("parent", "parenting"),
               c("phylogenies", "phylogeny"),
               c("phylogenetic", "phylogeny"),
               c("phylogenomic", "phylogeny"),
               c("pollinators", "pollination"),
               c("protected", "protect"),
               c("protective", "protect"),
               c("rainfall", "rain"),
               c("reconstructing", "reconstruction"),
               c("regulatory", "regulation"),
               c("regulates", "regulation"),
               c("relation", "relationship"),
               c("reproductive", "reproduction"),
               c("restored", "restoration"),
               c("robustness", "robust"),
               c("scientific", "science"),
               c("scientist", "science"),
               c("sexy", "sexual"),
               c("simulated", "simulation"),
               c("societies", "society"),
               c("social", "society"),
               c("socio", "society"),
               c("space", "spatial"),
               c("spacio", "spatial"),
               c("stabilize", "stability"),
               c("stable", "stability"),
               c("stories", "story"),
               c("strategic", "strategy"),
               c("strategies", "strategy"),
               c("structured", "structure"),
               c("structuring", "structure"),
               c("studies", "study"),
               c("studing", "study"),
               c("sustainable", "sustainability"),
               c("theories", "theory"),
               c("theoretical", "theory"),
               c("threatened", "threat"),
               c("tropical", "tropic"),
               c("vision", "visual")
               )
lemma <- as.data.frame(lemma)

for (i in 1:dim(lemma)[1]){
  text$word[text$word == lemma[i,1]] <- lemma[i,2]
}

WORDS - all data

Number of words per gender and academic position

table(text$gender) %>% kable()

Var1	Freq
F	10905
M	13429

table(text$position_cat ,text$gender)%>% kable()

	F	M
others	260	137
postdoc	2806	2555
professor	1977	5078
student	5771	5648

20 more common workds

text %>%
  count(word, sort = TRUE) %>% 
  top_n(20,n)%>%
  kable()

word	n
species	388
ecology	196
forest	187
study	161
model	159
evolution	140
environment	138
landscape	126
population	126
diversity	117
plant	103
community	101
male	101
different	97
nature	94
patterns	89
areas	86
interaction	84
present	84
animal	82
use	82

Word cloud

All words

textplot_wordcloud(x=dfm(tokens(text$word)))

By gender. Purple female, yellow male

par(mfrow=c(1,2))
textplot_wordcloud(x=dfm(tokens(text$word[text$gender=="F"])),
                   col="#6D57CF")
par(new=T)
textplot_wordcloud(x=dfm(tokens(text$word[text$gender=="M"])),
                   col="#FCA532")

Word frequencies by gender

props <- text %>%
  count(gender, word) %>%
  group_by(gender) %>%
  mutate(proportion = n / sum(n)) %>% 
  pivot_wider(names_from = gender, values_from = c(proportion,n)) %>%
  mutate(abs.dif.p = abs(proportion_F-proportion_M),
         rel.dif.p = pmax(proportion_F, proportion_M)/
           pmin(proportion_F, proportion_M)) %>%
  arrange(desc(abs.dif.p))
props$label <- NA
props$label[1:20] <- props$word[1:20]

Words that are close to the dashed line have similar frequencies in both genders. Words that are far from the line are words that are found more in one set of texts than another.

Legend: absolute differences in the frequency of the word by males and females. Differences above 0.3% are also indicated in text.

Correlation of word frequeency use between gender:

cor.test(props$proportion_F, props$proportion_M)

## 
##  Pearson's product-moment correlation
## 
## data:  props$proportion_F and props$proportion_M
## t = 70.674, df = 1650, p-value < 2.2e-16
## alternative hypothesis: true correlation is not equal to 0
## 95 percent confidence interval:
##  0.8544996 0.8784931
## sample estimates:
##      cor 
## 0.866998

Highly correlated -> it means they tend to use the same frequency of main word

20 words with the largest differences in frequency

prop2 <- props %>% filter(!is.na(label)) %>%
  arrange(desc(proportion_F), desc(proportion_M)) %>%
  mutate(ntot = n_F + n_M) %>%
  mutate(word = fct_reorder(word,(ntot),max),
         proportion_F = proportion_F*-1) %>%
  pivot_longer(2:3,names_to = "gender", values_to ="proportion")

ggplot(prop2, aes(x=proportion, y=word,fill=gender)) +
  geom_col()+ ylab("") + xlab("Proportion")+
  scale_fill_manual(name="gender", values=c("#6D57CF","#FCA532"),
                    labels=c("F", "M"))+
   geom_vline(xintercept = c(-0.02,-0.01,0,0.01,0.02),
              linetype="dotted",
             col="darkgray") +
  scale_x_continuous(breaks=c(-0.02,-0.01,0,0.01,0.02),
                     labels = c(0.02,0.01,0,0.01,0.02))

TF IDF

The statistic tf-idf is intended to measure how important a word is to a document in a collection (or corpus) of documents, for example, to one novel in a collection of novels or to one website in a collection of websites.

Calculating tf-idf attempts to find the words that are important (i.e., common) in a text, but not too common. Let’s do that now.

text_id <- text %>% count(gender, word) %>% 
  bind_tf_idf(word, gender, n) %>%
  arrange(desc(tf_idf))

10 “exclusive” words for each group

text_id$word <- as.factor(text_id$word)
text_id %>%
  group_by(gender) %>% 
  arrange(desc(tf_idf)) %>% 
  top_n(10, tf_idf) %>%  
  ggplot(aes(x = tf_idf, y = reorder(word, tf_idf), fill = gender)) +
  geom_col(show.legend = FALSE) +
  labs(x = NULL, y = "tf-idf") +
  facet_wrap(~gender, scales = "free") +
  theme_minimal()

WORDS - professors only data

textP <- text %>% filter(position_cat == "professor")

table(textP$gender)

## 
##    F    M 
## 1977 5078

par(mfrow=c(1,2))
textplot_wordcloud(x=dfm(tokens(textP$word[textP$gender=="F"])),
                   col="#6D57CF")
par(new=T)
textplot_wordcloud(x=dfm(tokens(textP$word[textP$gender=="M"])),
                   col="#FCA532")

Mean number of words by abstract

textP %>% count(id,gender) %>%
  ggplot(aes(x=gender, y=n)) + 
  geom_violin() + geom_boxplot(width=0.2)+
  ggbeeswarm::geom_quasirandom(size=3, shape=21)

20 most commmon words

textP %>%
  count(word, sort = TRUE) %>% 
  top_n(20,n)%>%
  kable()

word	n
species	90
ecology	71
evolution	54
population	50
environment	47
study	45
plant	43
model	39
diversity	37
nature	36
ecosystem	33
pollination	32
research	31
society	30
biology	29
science	29
interaction	27
amphibian	25
may	25
community	24

Words Frequency by gender

propsP <- textP %>%
    count(gender, word) %>%
  group_by(gender) %>%
  mutate(proportion = n / sum(n)) %>% 
  pivot_wider(names_from = gender, values_from = c(proportion,n)) %>%
  mutate(abs.dif.p = abs(proportion_F-proportion_M),
         rel.dif.p = pmax(proportion_F, proportion_M)/
           pmin(proportion_F, proportion_M)) %>%
  arrange(desc(abs.dif.p))
propsP$label <- NA
propsP$label[1:20] <- propsP$word[1:20]

ggplot(propsP, aes(x=proportion_M,, y=proportion_F,
       color=abs.dif.p)) + 
  geom_abline(color = "gray40", lty = 2) +
  geom_point(size=1, alpha=0.1)+
 # geom_jitter(size=2.5, alpha=0.02)+
  geom_text(aes(label=word), check_overlap = T,vjust=1,hjust=0.32)+
  #geom_text_repel(aes(label=label), size=3.2)+
  #xlim(0,1)+ylim(0,0.2)+
  scale_x_log10(name="Male most used words", limits=c(0.0003,0.02),
                labels = percent_format()) +
  scale_y_log10(name="Female most used words",limits=c(0.0003,0.02),
                labels = percent_format()) +
  scale_color_gradient(name="Absolute \n difference",low = "blue", high = "red",
                       labels=percent_format()
                       )  +
  theme(legend.justification = c(1, -0.1), legend.position = c(1, 0))

 # geom_smooth(method="lm")
ggsave("figures/FIG_S4_wordFreq_Prof.jpg", height = 7, width=9)

Words that are close to the dashed line in these plots have similar frequencies in both genders. Words that are far from the line are words that are found more in one set of texts than another.

Legend: absolute differences in the frequency of the word by males and females.

Labels for the 20 words with largest differences in frequency.

Correlation of word frequeency use between gender:

cor.test(propsP$proportion_F, propsP$proportion_M)

## 
##  Pearson's product-moment correlation
## 
## data:  propsP$proportion_F and propsP$proportion_M
## t = 20.205, df = 518, p-value < 2.2e-16
## alternative hypothesis: true correlation is not equal to 0
## 95 percent confidence interval:
##  0.6128988 0.7093872
## sample estimates:
##       cor 
## 0.6638971

20 words with the largest differences in frequency

propP2 <- propsP %>% filter(!is.na(label)) %>%
  arrange(desc(proportion_F), desc(proportion_M)) %>%
  mutate(ntot = n_F + n_M) %>%
  mutate(word = fct_reorder(word,(ntot),max),
         proportion_F = proportion_F*-1) %>%
  pivot_longer(2:3,names_to = "gender", values_to ="proportion")

ggplot(propP2, aes(x=proportion, y=word,fill=gender)) +
  geom_col()+ ylab("") + xlab("Proportion")+
  scale_fill_manual(name="gender", values=c("#6D57CF","#FCA532"),
                    labels=c("F", "M"))+
   geom_vline(xintercept = c(-0.02,-0.01,0,0.01,0.02),
              linetype="dotted",
             col="darkgray") +
  scale_x_continuous(breaks=c(-0.02,-0.01,0,0.01,0.02),
                     labels = c(0.02,0.01,0,0.01,0.02))

TF IDF

text_idP <- textP %>% count(gender, word) %>% 
  bind_tf_idf(word, gender, n) %>%
  arrange(desc(tf_idf))

10 “exclusive” words for each group

text_idP$word <- as.factor(text_idP$word)
text_idP %>%
  group_by(gender) %>% 
  arrange(desc(tf_idf)) %>% 
  top_n(10, tf_idf) %>%  
  ggplot(aes(x = tf_idf, y = reorder(word, tf_idf), fill = gender)) +
  geom_col(show.legend = FALSE) +
  labs(x = NULL, y = "tf-idf") +
  facet_wrap(~gender, scales = "free") +
  theme_minimal()

Topic model - all data

LDA - latent Dirichlet allocation method for fiting topic model

It treats each document as a mixture of topics, and each topic as a mixture of words. This allows documents to “overlap” each other in terms of content, rather than being separated into discrete groups, in a way that mirrors typical use of natural lan‐ guage.

Every document is a mixture of topics

Every topic is a mixture of words

matext <- text %>% count(id, gender, word) %>% mutate(id = paste(id, gender, sep="_")) %>%
  select(-gender) %>%
                cast_dtm(term=word,document=id,value=n)

Choosing number of topics: comparing AIC

ap_lda2 <- LDA(matext, k = 2, control = list(seed = 1234))
ap_lda3 <- LDA(matext, k = 3, control = list(seed = 1234))
ap_lda4 <- LDA(matext, k = 4, control = list(seed = 1234))
ap_lda5 <- LDA(matext, k = 5, control = list(seed = 1234))
ap_lda10<- LDA(matext, k = 10, control = list(seed = 1234))
ap_lda20 <- LDA(matext, k = 20, control = list(seed = 1234))
bbmle::AICtab(ap_lda2, ap_lda3, ap_lda4,ap_lda5, ap_lda10,ap_lda20,
              base=T)

##          AIC      dAIC     df   
## ap_lda2  375926.2      0.0 9877 
## ap_lda3  378052.8   2126.6 14815
## ap_lda4  382295.0   6368.9 19753
## ap_lda5  387617.8  11691.7 24691
## ap_lda10 419995.2  44069.0 49381
## ap_lda20 498167.4 122241.3 98761

two-topics model seems the most plausible model

Word-topic probabilities

10 words with the largest probabilities for each group

ap_topics <- tidy(ap_lda2, matrix = "beta")
ap_top_terms <- ap_topics %>%
      group_by(topic) %>%
      top_n(10, beta) %>%
      ungroup() %>%
      arrange(topic, -beta)
ap_top_terms %>%
mutate(term = reorder(term, beta)) %>% 
  ggplot(aes(term, beta, fill = factor(topic))) + 
  geom_col(show.legend = FALSE) +
facet_wrap(~ topic, scales = "free") + coord_flip()

words with the greates difference in Beta between topics

beta_spread <- ap_topics %>%
  mutate(topic = paste0("topic", topic)) %>%
  spread(topic, beta) %>%
  filter(topic1 > .001 | topic2 > .001) %>%
  mutate(log_ratio = log2(topic2 / topic1))
beta_spread

## # A tibble: 254 × 4
##    term           topic1   topic2 log_ratio
##    <chr>           <dbl>    <dbl>     <dbl>
##  1 abundance   0.00155   0.000501   -1.63  
##  2 action      0.000531  0.00134     1.34  
##  3 activities  0.000459  0.00116     1.33  
##  4 adaptation  0.000769  0.00222     1.53  
##  5 addition    0.00100   0.00106     0.0791
##  6 advance     0.0000853 0.00115     3.75  
##  7 affect      0.000752  0.00117     0.638 
##  8 agriculture 0.00106   0.00135     0.355 
##  9 also        0.00283   0.00275    -0.0386
## 10 amazon      0.000871  0.00130     0.577 
## # ℹ 244 more rows

beta_spread %>%  
  arrange(log_ratio) %>% slice(c(1:5,260:264)) %>%
  ggplot(aes(fct_reorder(term,log_ratio,min), log_ratio)) + 
  geom_col(show.legend = FALSE) + coord_flip() +
  ylab("Log2 ration of beta in topic 2/topic 1") + xlab("Word")

Document-topic probabilities - classifying the abstracts

and comparing the two groups by gender (if there is a difference in frequency)

ap_documents <- tidy(ap_lda2, matrix = "gamma")
classifi <- ap_documents %>% mutate(gender = substr(document, nchar(document), nchar(document))) %>%
  group_by(document,gender) %>%
      top_n(1, gamma) 

table(classifi$gender, classifi$topic)

##    
##      1  2
##   F 76 57
##   M 93 84

classifi %>% tabyl(gender, topic) %>% adorn_percentages() %>%
  adorn_pct_formatting(digits = 0) %>%
  adorn_ns() %>% kable()

gender	1	2
F	57% (76)	43% (57)
M	53% (93)	47% (84)

classifi %>%
 # mutate(title = reorder(title, gamma * topic)) %>%
  ggplot(aes(as.character(topic), gamma)) +
  geom_boxplot() +
  facet_wrap(~ gender)

Chi-square test

chisq.test(classifi$gender, classifi$topic)

## 
##  Pearson's Chi-squared test with Yates' continuity correction
## 
## data:  classifi$gender and classifi$topic
## X-squared = 0.47591, df = 1, p-value = 0.4903

Topic model - Professors only

matextP <- textP %>% 
  count(id, gender, word) %>% mutate(id = paste(id, gender, sep="_")) %>%
  select(-gender) %>%
                cast_dtm(term=word,document=id,value=n)

ap_lda2P <- LDA(matextP, k = 2, control = list(seed = 1234))
ap_lda3P <- LDA(matextP, k = 3, control = list(seed = 1234))
ap_lda4P <- LDA(matextP, k = 4, control = list(seed = 1234))
bbmle::AICtab(ap_lda2P, ap_lda3P, ap_lda4P,base=T)

##          AIC      dAIC     df  
## ap_lda2P 107292.8      0.0 4947
## ap_lda3P 108945.3   1652.4 7420
## ap_lda4P 111571.6   4278.8 9893

word-topic probabilities

ap_topicsP <- tidy(ap_lda2P, matrix = "beta")
ap_top_termsP <- ap_topicsP %>%
      group_by(topic) %>%
      top_n(10, beta) %>%
      ungroup() %>%
      arrange(topic, -beta)
ap_top_termsP %>%
mutate(term = reorder(term, beta)) %>% ggplot(aes(term, beta, fill = factor(topic))) + geom_col(show.legend = FALSE) +
facet_wrap(~ topic, scales = "free") + coord_flip()

words with the greates difference in Beta between topics

beta_spread <- ap_topicsP %>%
  mutate(topic = paste0("topic", topic)) %>%
  spread(topic, beta) %>%
  filter(topic1 > .001 | topic2 > .001) %>%
  mutate(log_ratio = log2(topic2 / topic1))
beta_spread

## # A tibble: 348 × 4
##    term         topic1   topic2 log_ratio
##    <chr>         <dbl>    <dbl>     <dbl>
##  1 abundance  0.00132  5.08e-44  -134.   
##  2 accepted   0.00132  1.47e-37  -113.   
##  3 across     0.000791 1.53e- 3     0.953
##  4 action     0.00153  3.66e- 4    -2.06 
##  5 activities 0.00177  7.01e- 4    -1.34 
##  6 adaptation 0.00161  3.95e- 3     1.30 
##  7 addition   0.00106  9.98e-10   -20.0  
##  8 address    0.000528 1.23e- 3     1.22 
##  9 advance    0.000538 1.52e- 3     1.50 
## 10 affect     0.00201  4.20e- 4    -2.26 
## # ℹ 338 more rows

beta_spread %>%  
  arrange(log_ratio) %>% slice(c(1:5,260:264)) %>%
  ggplot(aes(fct_reorder(term,log_ratio,min), log_ratio)) + 
  geom_col(show.legend = FALSE) + coord_flip() +
  ylab("Log2 ration of beta in topic 2/topic 1") + xlab("Word")

Document-topic probabilities

ap_documentsP <- tidy(ap_lda2P, matrix = "gamma")
classifiP <- ap_documentsP %>% mutate(gender = substr(document, nchar(document), nchar(document))) %>%
  group_by(document,gender) %>%
      top_n(1, gamma) 

table(classifiP$gender, classifiP$topic)

##    
##      1  2
##   F 12 12
##   M 39 33

library(janitor)
classifiP %>% tabyl(gender, topic) %>% adorn_percentages() %>% 
  adorn_pct_formatting(digits = 0) %>%
  adorn_ns() %>% kable()

gender	1	2
F	50% (12)	50% (12)
M	54% (39)	46% (33)

classifiP %>%
 # mutate(title = reorder(title, gamma * topic)) %>%
  ggplot(aes(as.character(topic), gamma)) +
  geom_boxplot() +
  geom_violin()+
  facet_wrap(~ gender)

Chi-square test

chisq.test(classifiP$gender, classifiP$topic)

## 
##  Pearson's Chi-squared test with Yates' continuity correction
## 
## data:  classifiP$gender and classifiP$topic
## X-squared = 0.013943, df = 1, p-value = 0.906